We Audited Our Own Site FirstOceanads.org Website Self-Audit Report
More and more people skip Google and ask AI directly: "Which web design agencies in Taichung are reliable?" If you are not in the AI's answer, those customers never hear about you.
What we do is called GEO: we look at how AI tools like ChatGPT, Gemini and Perplexity answer questions about your industry, find out why they leave you out, then fix your website and outside sources until AI can read them and is willing to cite them.
This is the full audit we ran on our own site. The result is not pretty: across 160 answers from five AIs, we were mentioned only once.
Audit and sampling, side by side.
- 01
Check readability
First, see whether machines can read the site and whether page content is complete.
- 02
Review the actual scores
Compare the six dimensions and the page distribution to find what needs work.
- 03
Compare against AI answers
Show the sampling results together with their limits, then decide the next step.
How these AIs answer your customers can be optimized
Each AI finds information differently: some rely on memory, others search the web in real time. We sample each one and adjust for how it tends to cite sources.
International
- ChatGPTTested this round
- GeminiTested this round
- Google AI Overviews
- PerplexityTested this round
- GrokTested this round
- Claude
Mainland China
- DeepSeekTested this round
- Doubao
- Kimi
- Zhipu GLM
- MiniMax
- Baidu AI Search
Where AI answers come from, and where we step in
- Customer asks "Who should I hire to build a website in Taichung?" Build a bank of questions your customers actually ask
- AI looks for information From memory, or a live web search Make sure AI can crawl and understand your site
- Picks sources to cite Review platforms, rankings, media, official sites Get your information onto the platforms AI cites most
- Writes the answer Are you in it, what position, is it accurate Re-ask each AI and verify with numbers
marks our work
Can AI read our website?
This step is the ticket in. Everything passed except one item: hreflang.
| Check | Result | Verdict |
|---|---|---|
| Pages accessible | 25 / 25 | Pass |
| sitemap.xml | 119 URLs listed | Pass |
| llms.txt | Valid format | Pass |
| robots.txt blocks AI | None | Pass |
| AI crawler live test | GPTBot, ClaudeBot, PerplexityBot, Bytespider all returned 200 | Pass |
| Chinese/English page mapping (hreflang) | 7 / 25 | To fix |
Scores on six dimensions
Site average: 79.0. The weakest is query match: title wording does not match how customers ask, on 21 pages.
| Dimension | Avg / Max | Score rate | Distribution |
|---|---|---|---|
| Crawlability | 15.00 / 15 | 100% | |
| Content length | 13.86 / 15 | 92% | |
| Structure | 17.92 / 20 | 90% | |
| Extractable blocks | 20.16 / 25 | 81% | |
| Authority signals | 9.44 / 15 | 63% | |
| Query match | 2.62 / 10 | 26% |
In empirical research, query match is the single most influential factor (correlation 0.432), ahead of semantic similarity (0.356) and content quality (0.292).
Page grades across three audit rounds
More A grades, and the first C grades. What pulled scores down: essay-style articles written in the last two months.
| Page | Aug | Nov | Reason |
|---|---|---|---|
| en/case-esim.html | 71.2 | 83.3 | Added a comparison section |
| en/case-yuxiang.html | 72.5 | 83.3 | Added a comparison section |
| en/case-gpnet.html | 74.0 | 85.8 | Added a comparison section |
| blog/article-92.html | — | 59.1 | New article; no passage can be cited on its own |
| blog/article-96.html | — | 60.2 | New article; lacks definitions, numbers and steps |
Where we lack passages AI can lift directly
AI cites passages, not whole pages. The more of these five passage types a page lacks, the worse it does.
| Passage type | Pages missing | Distribution | Effect size |
|---|---|---|---|
| Step-by-step instructions | 10 | +41.2% | |
| FAQ | 6 | Helps Q&A recall | |
| Definitions | 5 | +57.3% | |
| Numeric facts | 3 | +61.6% | |
| Comparisons | 1 | +55.3% |
How five AIs actually answered
32 recommendation questions that never include our name, to see whether AI thinks of us on its own.
| AI | Web search | Mentioned us | Cited our site |
|---|---|---|---|
| Perplexity | Yes | 1 / 32 | 3 / 32 |
| Gemini | No | 0 / 32 | 0 / 32 |
| Grok | No | 0 / 32 | 0 / 32 |
| ChatGPT | No | 0 / 32 | 0 / 32 |
| DeepSeek | No | 0 / 32 | 0 / 32 |
| Brand | Mentions | Distribution |
|---|---|---|
| Shopify | 47 | |
| SHOPLINE | 31 | |
| CYBERBIZ | 24 | |
| WordPress | 22 | |
| 91APP | 11 | |
| OceanAds | 1 |
Asked by name: "What does OceanAds (海水不可斗量) do?"
| AI | Named us | Was it accurate? |
|---|---|---|
| Perplexity | 5 / 5 | 4 correct, based on our website and PRO360 reviews; 1 treated the name as a Chinese idiom |
| Gemini | 5 / 5 | 1 answer described us as a curry restaurant company in Taichung |
| Grok | 2 / 5 | Most answers found no information and suggested checking the business registry |
| ChatGPT | 4 / 5 | Said plainly it could not find us and suggested the business registry; made nothing up |
| DeepSeek | 3 / 5 | 4 said plainly they could not verify; 1 described us as a saltwater aquarium shop in Taichung |
"[The name] is not a traditional large enterprise, but a restaurant company that has become quite well known in Taiwan in recent years (especially in Taichung), mainly operating aged-curry and related food brands."
Gemini answer (translated) · 2026-11-13 · question q217
Naming us is not the same as getting us right: Gemini said we sell curry, DeepSeek said we are an aquarium shop. When an AI without web access finds no data, it fills the gap from the literal meaning of the name ("seawater").
Sampling conditions
| Platforms | ChatGPT, Gemini, Perplexity Sonar (via OpenRouter), Grok (via kie.ai), DeepSeek (official API) |
|---|---|
| Endpoint | API, evidence grade B (reproducible) |
| Web search | Perplexity on; off for the other four |
| Questions | 37 total: 32 recommendation, 5 brand verification |
| Calculation | Web-search and no-search results listed separately; no overall average |
Most of what AI cites is not on your website
Your website makes sure AI describes you correctly. To get recommended, you need to appear on the platforms AI cites most.
| Source type | Share | Distribution |
|---|---|---|
| Content platforms | 16.4% | |
| News media | 13.6% | |
| Recommendation & ranking sites | 9.1% | |
| Brand websites | 1.37% |
| Domain | Citations | Distribution |
|---|---|---|
| clutch.co | 10 | |
| da-vinci.com.tw | 7 | |
| vocus.cc | 7 | |
| goodfirms.co | 6 | |
| frankknow.com | 6 | |
| reddit.com | 5 |
The table on the right is the work order for next round's external source building: review directories (clutch.co, goodfirms.co), content platforms (vocus.cc), and peer recommendation articles.
What we will do next
No P0 (ticket-in) issues. Of last round's 9 tickets: 0 passed, 5 fell short, 4 await manual confirmation.
| Priority | Action | Pages affected |
|---|---|---|
| P1 | Rewrite titles in the words customers actually use | 21 |
| P1 | Add "step-by-step" passages | 10 |
| P1 | Add "FAQ" passages | 6 |
| P1 | Add "definition" passages | 5 |
| P1 | Add "numeric fact" passages | 3 |
| P1 | Rework essay articles into 2–3 passages that can be cited on their own | 2 |
| P1 | Turn key points into lists; split into 6–10 H2 sections | 2 |
| P1 | Align FAQPage structured data with on-page Q&A | 2 |
| P1 | Complete hreflang for Chinese/English pages | Site-wide |
| P1 | Add a visible last-updated date to the homepage | 1 |
| P2 | Add citations to external sources | 25 |
The same process, applied to your website
Five services. Take the full package, or pick just one or two.
| Service | Who it's for | Deliverables | Done when |
|---|---|---|---|
| GEO Audit Report | You want to know where you stand in AI answers first | A report in the same format as this page, plus one online walkthrough | Report delivered and walked through with you |
| Technical Page Fixes | Any website; JS-framework sites first | robots.txt, llms.txt, JSON-LD, before/after score comparison | Full 15 points on crawlability after a re-crawl |
| Content Engineering | Technical issues fixed, now aiming for AI answers (about 3–6 months) | A content outline for each target question; ghostwriting available | Extractable-block and query-match scores rise on re-test |
| External Source Building | Content is in place, but AI still doesn't mention you | A source list ranked by sampling results, plus a unified brand fact sheet | Publication and review outcomes confirmed by hand |
| AI Answer Monitoring | You need to report results to your boss or clients | Diff report each round, cited-domain ranking, raw sampling files | Ongoing; we suggest one round every 2–4 weeks |
- The audit is freeEngagements are quoted by site size, with no long-term contract.
- Accounts in your nameWe work through invited access; reports and raw files belong to you.
- We confirm before we actBefore changing your site we list what we'll do, and only proceed once you agree.
What we don't promise
- No guarantee of citationGEO raises the probability of being cited. No one can guarantee it.
- No passing correlation off as causationCausation needs a baseline, control question sets and competitor comparisons. We always state whether we have them.
- No record, no sampleEvery entry keeps the original answer text and the sampling environment.
- If we can't find it, we mark it unconfirmedWe never fill in brand data with guesswork.
Audit date: November 13, 2026. The previous two rounds were August 4 and October 11. For the scoring method, see how we analyze.
FAQ
How is GEO different from SEO?
SEO is about rankings and clicks, and each result holds one position. A generative engine gives a full paragraph: you might be written into it, listed fourth, described in the wrong industry, or not appear at all. The smallest unit GEO measures is an extractable fact block: a passage that can be lifted out on its own and still make sense.
Why does AI barely mention you when your site scores close to 80?
The two numbers measure different stages. The 79.0 score measures whether pages can be crawled and extracted, which is the condition for entering the candidate pool. Mention rate measures whether AI thinks of you when it answers. Without web access, ChatGPT, Gemini, Grok and DeepSeek rely only on model memory, and they have almost no impression of a small studio. Perplexity, which searches the web, listed us in its recommendations for one question, based on our website content. The site score is the ticket in; getting recommended takes external sources.
How often should a GEO audit be run?
We suggest page audits weekly and answer sampling every two weeks or monthly. Sampling has a cost and the metrics are noisy, so running it too often shows no signal.
Does GEO guarantee ChatGPT will cite me?
No one can guarantee that a given platform will cite a given page. GEO raises the probability of being retrieved, selected for the citation list and absorbed into the answer. By default we report attribution as observed correlation; claiming causation requires a baseline window, control question sets and competitor comparisons.
Want to see how your website looks in AI answers?
Send us your URL. We'll run the same method and put your six-dimension scores, extractable-block gaps and sampling results into a report in the same format. Decide after you've read it.